xref: /freebsd/sys/dev/hwpmc/hwpmc_amd.c (revision 146b30bad9f65a098e6c09ae93bb1da2ff59616d)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2003-2008 Joseph Koshy
5  * Copyright (c) 2007 The FreeBSD Foundation
6  * All rights reserved.
7  *
8  * Portions of this software were developed by A. Joseph Koshy under
9  * sponsorship from the FreeBSD Foundation and Google, Inc.
10  *
11  * Redistribution and use in source and binary forms, with or without
12  * modification, are permitted provided that the following conditions
13  * are met:
14  * 1. Redistributions of source code must retain the above copyright
15  *    notice, this list of conditions and the following disclaimer.
16  * 2. Redistributions in binary form must reproduce the above copyright
17  *    notice, this list of conditions and the following disclaimer in the
18  *    documentation and/or other materials provided with the distribution.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
21  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
24  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30  * SUCH DAMAGE.
31  */
32 
33 /* Support for the AMD K8 and later processors */
34 
35 #include <sys/param.h>
36 #include <sys/lock.h>
37 #include <sys/malloc.h>
38 #include <sys/mutex.h>
39 #include <sys/pcpu.h>
40 #include <sys/pmc.h>
41 #include <sys/pmckern.h>
42 #include <sys/smp.h>
43 #include <sys/sysctl.h>
44 #include <sys/systm.h>
45 
46 #include <machine/cpu.h>
47 #include <machine/cpufunc.h>
48 #include <machine/md_var.h>
49 #include <machine/specialreg.h>
50 
51 #define	OVERFLOW_WAIT_COUNT	50
52 
53 DPCPU_DEFINE_STATIC(uint32_t, nmi_counter);
54 
55 /* AMD K8 PMCs */
56 struct amd_descr {
57 	struct pmc_descr pm_descr;   /* "base class" */
58 	uint32_t	pm_evsel;    /* address of EVSEL register */
59 	uint32_t	pm_perfctr;  /* address of PERFCTR register */
60 	enum sub_class	pm_subclass; /* register subclass */
61 };
62 
63 static int amd_npmcs;
64 static int amd_core_npmcs, amd_l3_npmcs, amd_df_npmcs;
65 static struct amd_descr amd_pmcdesc[AMD_NPMCS_MAX];
66 struct amd_event_code_map {
67 	enum pmc_event	pe_ev;	 /* enum value */
68 	uint16_t	pe_code; /* encoded event mask */
69 	uint8_t		pe_mask; /* bits allowed in unit mask */
70 };
71 
72 const struct amd_event_code_map amd_event_codes[] = {
73 	{ PMC_EV_K8_FP_DISPATCHED_FPU_OPS,		0x00, 0x3F },
74 	{ PMC_EV_K8_FP_CYCLES_WITH_NO_FPU_OPS_RETIRED,	0x01, 0x00 },
75 	{ PMC_EV_K8_FP_DISPATCHED_FPU_FAST_FLAG_OPS,	0x02, 0x00 },
76 
77 	{ PMC_EV_K8_LS_SEGMENT_REGISTER_LOAD, 		0x20, 0x7F },
78 	{ PMC_EV_K8_LS_MICROARCHITECTURAL_RESYNC_BY_SELF_MODIFYING_CODE,
79 	  						0x21, 0x00 },
80 	{ PMC_EV_K8_LS_MICROARCHITECTURAL_RESYNC_BY_SNOOP, 0x22, 0x00 },
81 	{ PMC_EV_K8_LS_BUFFER2_FULL,			0x23, 0x00 },
82 	{ PMC_EV_K8_LS_LOCKED_OPERATION,		0x24, 0x07 },
83 	{ PMC_EV_K8_LS_MICROARCHITECTURAL_LATE_CANCEL,	0x25, 0x00 },
84 	{ PMC_EV_K8_LS_RETIRED_CFLUSH_INSTRUCTIONS,	0x26, 0x00 },
85 	{ PMC_EV_K8_LS_RETIRED_CPUID_INSTRUCTIONS,	0x27, 0x00 },
86 
87 	{ PMC_EV_K8_DC_ACCESS,				0x40, 0x00 },
88 	{ PMC_EV_K8_DC_MISS,				0x41, 0x00 },
89 	{ PMC_EV_K8_DC_REFILL_FROM_L2,			0x42, 0x1F },
90 	{ PMC_EV_K8_DC_REFILL_FROM_SYSTEM,		0x43, 0x1F },
91 	{ PMC_EV_K8_DC_COPYBACK,			0x44, 0x1F },
92 	{ PMC_EV_K8_DC_L1_DTLB_MISS_AND_L2_DTLB_HIT,	0x45, 0x00 },
93 	{ PMC_EV_K8_DC_L1_DTLB_MISS_AND_L2_DTLB_MISS,	0x46, 0x00 },
94 	{ PMC_EV_K8_DC_MISALIGNED_DATA_REFERENCE,	0x47, 0x00 },
95 	{ PMC_EV_K8_DC_MICROARCHITECTURAL_LATE_CANCEL,	0x48, 0x00 },
96 	{ PMC_EV_K8_DC_MICROARCHITECTURAL_EARLY_CANCEL, 0x49, 0x00 },
97 	{ PMC_EV_K8_DC_ONE_BIT_ECC_ERROR,		0x4A, 0x03 },
98 	{ PMC_EV_K8_DC_DISPATCHED_PREFETCH_INSTRUCTIONS, 0x4B, 0x07 },
99 	{ PMC_EV_K8_DC_DCACHE_ACCESSES_BY_LOCKS,	0x4C, 0x03 },
100 
101 	{ PMC_EV_K8_BU_CPU_CLK_UNHALTED,		0x76, 0x00 },
102 	{ PMC_EV_K8_BU_INTERNAL_L2_REQUEST,		0x7D, 0x1F },
103 	{ PMC_EV_K8_BU_FILL_REQUEST_L2_MISS,		0x7E, 0x07 },
104 	{ PMC_EV_K8_BU_FILL_INTO_L2,			0x7F, 0x03 },
105 
106 	{ PMC_EV_K8_IC_FETCH,				0x80, 0x00 },
107 	{ PMC_EV_K8_IC_MISS,				0x81, 0x00 },
108 	{ PMC_EV_K8_IC_REFILL_FROM_L2,			0x82, 0x00 },
109 	{ PMC_EV_K8_IC_REFILL_FROM_SYSTEM,		0x83, 0x00 },
110 	{ PMC_EV_K8_IC_L1_ITLB_MISS_AND_L2_ITLB_HIT,	0x84, 0x00 },
111 	{ PMC_EV_K8_IC_L1_ITLB_MISS_AND_L2_ITLB_MISS,	0x85, 0x00 },
112 	{ PMC_EV_K8_IC_MICROARCHITECTURAL_RESYNC_BY_SNOOP, 0x86, 0x00 },
113 	{ PMC_EV_K8_IC_INSTRUCTION_FETCH_STALL,		0x87, 0x00 },
114 	{ PMC_EV_K8_IC_RETURN_STACK_HIT,		0x88, 0x00 },
115 	{ PMC_EV_K8_IC_RETURN_STACK_OVERFLOW,		0x89, 0x00 },
116 
117 	{ PMC_EV_K8_FR_RETIRED_X86_INSTRUCTIONS,	0xC0, 0x00 },
118 	{ PMC_EV_K8_FR_RETIRED_UOPS,			0xC1, 0x00 },
119 	{ PMC_EV_K8_FR_RETIRED_BRANCHES,		0xC2, 0x00 },
120 	{ PMC_EV_K8_FR_RETIRED_BRANCHES_MISPREDICTED,	0xC3, 0x00 },
121 	{ PMC_EV_K8_FR_RETIRED_TAKEN_BRANCHES,		0xC4, 0x00 },
122 	{ PMC_EV_K8_FR_RETIRED_TAKEN_BRANCHES_MISPREDICTED, 0xC5, 0x00 },
123 	{ PMC_EV_K8_FR_RETIRED_FAR_CONTROL_TRANSFERS,	0xC6, 0x00 },
124 	{ PMC_EV_K8_FR_RETIRED_RESYNCS,			0xC7, 0x00 },
125 	{ PMC_EV_K8_FR_RETIRED_NEAR_RETURNS,		0xC8, 0x00 },
126 	{ PMC_EV_K8_FR_RETIRED_NEAR_RETURNS_MISPREDICTED, 0xC9, 0x00 },
127 	{ PMC_EV_K8_FR_RETIRED_TAKEN_BRANCHES_MISPREDICTED_BY_ADDR_MISCOMPARE,
128 							0xCA, 0x00 },
129 	{ PMC_EV_K8_FR_RETIRED_FPU_INSTRUCTIONS,	0xCB, 0x0F },
130 	{ PMC_EV_K8_FR_RETIRED_FASTPATH_DOUBLE_OP_INSTRUCTIONS,
131 							0xCC, 0x07 },
132 	{ PMC_EV_K8_FR_INTERRUPTS_MASKED_CYCLES,	0xCD, 0x00 },
133 	{ PMC_EV_K8_FR_INTERRUPTS_MASKED_WHILE_PENDING_CYCLES, 0xCE, 0x00 },
134 	{ PMC_EV_K8_FR_TAKEN_HARDWARE_INTERRUPTS,	0xCF, 0x00 },
135 
136 	{ PMC_EV_K8_FR_DECODER_EMPTY,			0xD0, 0x00 },
137 	{ PMC_EV_K8_FR_DISPATCH_STALLS,			0xD1, 0x00 },
138 	{ PMC_EV_K8_FR_DISPATCH_STALL_FROM_BRANCH_ABORT_TO_RETIRE,
139 							0xD2, 0x00 },
140 	{ PMC_EV_K8_FR_DISPATCH_STALL_FOR_SERIALIZATION, 0xD3, 0x00 },
141 	{ PMC_EV_K8_FR_DISPATCH_STALL_FOR_SEGMENT_LOAD,	0xD4, 0x00 },
142 	{ PMC_EV_K8_FR_DISPATCH_STALL_WHEN_REORDER_BUFFER_IS_FULL,
143 							0xD5, 0x00 },
144 	{ PMC_EV_K8_FR_DISPATCH_STALL_WHEN_RESERVATION_STATIONS_ARE_FULL,
145 							0xD6, 0x00 },
146 	{ PMC_EV_K8_FR_DISPATCH_STALL_WHEN_FPU_IS_FULL,	0xD7, 0x00 },
147 	{ PMC_EV_K8_FR_DISPATCH_STALL_WHEN_LS_IS_FULL,	0xD8, 0x00 },
148 	{ PMC_EV_K8_FR_DISPATCH_STALL_WHEN_WAITING_FOR_ALL_TO_BE_QUIET,
149 							0xD9, 0x00 },
150 	{ PMC_EV_K8_FR_DISPATCH_STALL_WHEN_FAR_XFER_OR_RESYNC_BRANCH_PENDING,
151 							0xDA, 0x00 },
152 	{ PMC_EV_K8_FR_FPU_EXCEPTIONS,			0xDB, 0x0F },
153 	{ PMC_EV_K8_FR_NUMBER_OF_BREAKPOINTS_FOR_DR0,	0xDC, 0x00 },
154 	{ PMC_EV_K8_FR_NUMBER_OF_BREAKPOINTS_FOR_DR1,	0xDD, 0x00 },
155 	{ PMC_EV_K8_FR_NUMBER_OF_BREAKPOINTS_FOR_DR2,	0xDE, 0x00 },
156 	{ PMC_EV_K8_FR_NUMBER_OF_BREAKPOINTS_FOR_DR3,	0xDF, 0x00 },
157 
158 	{ PMC_EV_K8_NB_MEMORY_CONTROLLER_PAGE_ACCESS_EVENT, 0xE0, 0x7 },
159 	{ PMC_EV_K8_NB_MEMORY_CONTROLLER_PAGE_TABLE_OVERFLOW, 0xE1, 0x00 },
160 	{ PMC_EV_K8_NB_MEMORY_CONTROLLER_DRAM_COMMAND_SLOTS_MISSED,
161 							0xE2, 0x00 },
162 	{ PMC_EV_K8_NB_MEMORY_CONTROLLER_TURNAROUND,	0xE3, 0x07 },
163 	{ PMC_EV_K8_NB_MEMORY_CONTROLLER_BYPASS_SATURATION, 0xE4, 0x0F },
164 	{ PMC_EV_K8_NB_SIZED_COMMANDS,			0xEB, 0x7F },
165 	{ PMC_EV_K8_NB_PROBE_RESULT,			0xEC, 0x0F },
166 	{ PMC_EV_K8_NB_HT_BUS0_BANDWIDTH,		0xF6, 0x0F },
167 	{ PMC_EV_K8_NB_HT_BUS1_BANDWIDTH,		0xF7, 0x0F },
168 	{ PMC_EV_K8_NB_HT_BUS2_BANDWIDTH,		0xF8, 0x0F }
169 
170 };
171 
172 const int amd_event_codes_size = nitems(amd_event_codes);
173 
174 /*
175  * Per-processor information
176  */
177 struct amd_cpu {
178 	struct pmc_hw	pc_amdpmcs[AMD_NPMCS_MAX];
179 };
180 static struct amd_cpu **amd_pcpu;
181 
182 /* Populated by amd_init_policy(); PRECISERETIRE is OR-ed in per-allocation. */
183 static uint64_t amd_core_allowed_mask;
184 static uint64_t amd_l3_allowed_mask;
185 static uint64_t amd_df_allowed_mask;
186 
187 static uint64_t amd_core_extra_mask;
188 static uint64_t amd_l3_extra_mask;
189 static uint64_t amd_df_extra_mask;
190 
191 SYSCTL_DECL(_kern_hwpmc);
192 
193 SYSCTL_U64(_kern_hwpmc, OID_AUTO, amd_core_extra_mask, CTLFLAG_RDTUN,
194     &amd_core_extra_mask, 0,
195     "Extra allowed bits in AMD core PMU PERFEVTSEL (override; default 0)");
196 
197 SYSCTL_U64(_kern_hwpmc, OID_AUTO, amd_l3_extra_mask, CTLFLAG_RDTUN,
198     &amd_l3_extra_mask, 0,
199     "Extra allowed bits in AMD L3 PMU control (override; default 0)");
200 
201 SYSCTL_U64(_kern_hwpmc, OID_AUTO, amd_df_extra_mask, CTLFLAG_RDTUN,
202     &amd_df_extra_mask, 0,
203     "Extra allowed bits in AMD DF PMU control (override; default 0)");
204 
205 static void
amd_init_policy(void)206 amd_init_policy(void)
207 {
208 	int family;
209 
210 	family = CPUID_TO_FAMILY(cpu_id);
211 
212 	amd_core_allowed_mask = AMD_VALID_BITS;
213 
214 	amd_l3_allowed_mask = (family <= 0x17) ?
215 	    AMD_PMC_L3_FAMILY17_MASK : AMD_PMC_L3_FAMILY19_MASK;
216 
217 	amd_df_allowed_mask = (family <= 0x19) ?
218 	    AMD_PMC_DF_FAMILY17_MASK : AMD_PMC_DF_FAMILY1A_MASK;
219 }
220 
221 static uint64_t
amd_config_mask(enum sub_class subclass,uint64_t caps)222 amd_config_mask(enum sub_class subclass, uint64_t caps)
223 {
224 
225 	switch (subclass) {
226 	case PMC_AMD_SUB_CLASS_CORE:
227 		return (amd_core_allowed_mask | amd_core_extra_mask |
228 		    (((caps & PMC_CAP_PRECISE) != 0) ?
229 		    AMD_PMC_PRECISERETIRE : 0));
230 	case PMC_AMD_SUB_CLASS_L3_CACHE:
231 		return (amd_l3_allowed_mask | amd_l3_extra_mask);
232 	case PMC_AMD_SUB_CLASS_DATA_FABRIC:
233 		return (amd_df_allowed_mask | amd_df_extra_mask);
234 	default:
235 		return (0);
236 	}
237 }
238 
239 /*
240  * Read a PMC value from the MSR.
241  */
242 static int
amd_read_pmc(int cpu,int ri,struct pmc * pm,pmc_value_t * v)243 amd_read_pmc(int cpu, int ri, struct pmc *pm, pmc_value_t *v)
244 {
245 	const struct amd_descr *pd;
246 	pmc_value_t tmp;
247 	enum pmc_mode mode;
248 
249 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
250 	    ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
251 	KASSERT(ri >= 0 && ri < amd_npmcs,
252 	    ("[amd,%d] illegal row-index %d", __LINE__, ri));
253 	KASSERT(amd_pcpu[cpu],
254 	    ("[amd,%d] null per-cpu, cpu %d", __LINE__, cpu));
255 
256 	pd = &amd_pmcdesc[ri];
257 	mode = PMC_TO_MODE(pm);
258 
259 	PMCDBG2(MDP, REA, 1, "amd-read id=%d class=%d", ri,
260 	    pd->pm_descr.pd_class);
261 
262 	tmp = rdmsr(pd->pm_perfctr); /* RDMSR serializes */
263 	PMCDBG2(MDP, REA, 2, "amd-read (pre-munge) id=%d -> %jd", ri, tmp);
264 	if (PMC_IS_SAMPLING_MODE(mode)) {
265 		/*
266 		 * Clamp value to 0 if the counter just overflowed,
267 		 * otherwise the returned reload count would wrap to a
268 		 * huge value.
269 		 */
270 		if ((tmp & (1ULL << 47)) == 0)
271 			tmp = 0;
272 		else {
273 			/* Sign extend 48 bit value to 64 bits. */
274 			tmp = (pmc_value_t) ((int64_t)(tmp << 16) >> 16);
275 			tmp = AMD_PERFCTR_VALUE_TO_RELOAD_COUNT(tmp);
276 		}
277 	}
278 	*v = tmp;
279 
280 	PMCDBG2(MDP, REA, 2, "amd-read (post-munge) id=%d -> %jd", ri, *v);
281 
282 	return (0);
283 }
284 
285 /*
286  * Write a PMC MSR.
287  */
288 static int
amd_write_pmc(int cpu,int ri,struct pmc * pm,pmc_value_t v)289 amd_write_pmc(int cpu, int ri, struct pmc *pm, pmc_value_t v)
290 {
291 	const struct amd_descr *pd;
292 	enum pmc_mode mode;
293 
294 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
295 	    ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
296 	KASSERT(ri >= 0 && ri < amd_npmcs,
297 	    ("[amd,%d] illegal row-index %d", __LINE__, ri));
298 
299 	pd = &amd_pmcdesc[ri];
300 	mode = PMC_TO_MODE(pm);
301 
302 	/* use 2's complement of the count for sampling mode PMCs */
303 	if (PMC_IS_SAMPLING_MODE(mode))
304 		v = AMD_RELOAD_COUNT_TO_PERFCTR_VALUE(v);
305 
306 	PMCDBG3(MDP, WRI, 1, "amd-write cpu=%d ri=%d v=%jx", cpu, ri, v);
307 
308 	/* write the PMC value */
309 	wrmsr(pd->pm_perfctr, v);
310 	return (0);
311 }
312 
313 /*
314  * Configure hardware PMC according to the configuration recorded in 'pm'.
315  */
316 static int
amd_config_pmc(int cpu,int ri,struct pmc * pm)317 amd_config_pmc(int cpu, int ri, struct pmc *pm)
318 {
319 	struct pmc_hw *phw;
320 
321 	PMCDBG3(MDP, CFG, 1, "cpu=%d ri=%d pm=%p", cpu, ri, pm);
322 
323 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
324 	    ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
325 	KASSERT(ri >= 0 && ri < amd_npmcs,
326 	    ("[amd,%d] illegal row-index %d", __LINE__, ri));
327 
328 	phw = &amd_pcpu[cpu]->pc_amdpmcs[ri];
329 
330 	KASSERT(pm == NULL || phw->phw_pmc == NULL,
331 	    ("[amd,%d] pm=%p phw->pm=%p hwpmc not unconfigured",
332 		__LINE__, pm, phw->phw_pmc));
333 
334 	phw->phw_pmc = pm;
335 	return (0);
336 }
337 
338 /*
339  * Retrieve a configured PMC pointer from hardware state.
340  */
341 static int
amd_get_config(int cpu,int ri,struct pmc ** ppm)342 amd_get_config(int cpu, int ri, struct pmc **ppm)
343 {
344 	*ppm = amd_pcpu[cpu]->pc_amdpmcs[ri].phw_pmc;
345 	return (0);
346 }
347 
348 /*
349  * Machine-dependent actions taken during the context switch in of a
350  * thread.
351  */
352 static int
amd_switch_in(struct pmc_cpu * pc __pmcdbg_used,struct pmc_process * pp)353 amd_switch_in(struct pmc_cpu *pc __pmcdbg_used, struct pmc_process *pp)
354 {
355 	PMCDBG3(MDP, SWI, 1, "pc=%p pp=%p enable-msr=%d", pc, pp,
356 	    (pp->pp_flags & PMC_PP_ENABLE_MSR_ACCESS) != 0);
357 
358 	/* enable the RDPMC instruction if needed */
359 	if (pp->pp_flags & PMC_PP_ENABLE_MSR_ACCESS)
360 		load_cr4(rcr4() | CR4_PCE);
361 
362 	return (0);
363 }
364 
365 /*
366  * Machine-dependent actions taken during the context switch out of a
367  * thread.
368  */
369 static int
amd_switch_out(struct pmc_cpu * pc __pmcdbg_used,struct pmc_process * pp __pmcdbg_used)370 amd_switch_out(struct pmc_cpu *pc __pmcdbg_used,
371     struct pmc_process *pp __pmcdbg_used)
372 {
373 	PMCDBG3(MDP, SWO, 1, "pc=%p pp=%p enable-msr=%d", pc, pp, pp ?
374 	    (pp->pp_flags & PMC_PP_ENABLE_MSR_ACCESS) == 1 : 0);
375 
376 	/* always turn off the RDPMC instruction */
377 	load_cr4(rcr4() & ~CR4_PCE);
378 
379 	return (0);
380 }
381 
382 /*
383  * Check if a given PMC allocation is feasible.
384  */
385 static int
amd_allocate_pmc(int cpu __unused,int ri,struct pmc * pm,const struct pmc_op_pmcallocate * a)386 amd_allocate_pmc(int cpu __unused, int ri, struct pmc *pm,
387     const struct pmc_op_pmcallocate *a)
388 {
389 	const struct pmc_descr *pd;
390 	uint64_t allowed_unitmask, caps, config, unitmask;
391 	enum pmc_event pe;
392 	int i;
393 
394 	KASSERT(ri >= 0 && ri < amd_npmcs,
395 	    ("[amd,%d] illegal row index %d", __LINE__, ri));
396 
397 	pd = &amd_pmcdesc[ri].pm_descr;
398 
399 	/* check class match */
400 	if (pd->pd_class != a->pm_class)
401 		return (EINVAL);
402 
403 	if ((a->pm_flags & PMC_F_EV_PMU) == 0)
404 		return (EINVAL);
405 
406 	caps = pm->pm_caps;
407 
408 	if (((caps & PMC_CAP_PRECISE) != 0) &&
409 	    ((pd->pd_caps & PMC_CAP_PRECISE) == 0))
410 		return (EINVAL);
411 
412 	PMCDBG2(MDP, ALL, 1,"amd-allocate ri=%d caps=0x%x", ri, caps);
413 
414 	/* Validate sub-class. */
415 	if (amd_pmcdesc[ri].pm_subclass != a->pm_md.pm_amd.pm_amd_sub_class)
416 		return (EINVAL);
417 
418 	if (strlen(pmc_cpuid) != 0) {
419 		config = a->pm_md.pm_amd.pm_amd_config;
420 		if ((config & ~amd_config_mask(amd_pmcdesc[ri].pm_subclass,
421 		    caps)) != 0)
422 			return (EINVAL);
423 		pm->pm_md.pm_amd.pm_amd_evsel = config;
424 		PMCDBG2(MDP, ALL, 2, "amd-allocate ri=%d -> config=0x%jx",
425 		    ri, (uintmax_t)config);
426 		return (0);
427 	}
428 
429 	/*
430 	 * Everything below this is for supporting older processors.
431 	 */
432 	pe = a->pm_ev;
433 
434 	/* map ev to the correct event mask code */
435 	config = allowed_unitmask = 0;
436 	for (i = 0; i < amd_event_codes_size; i++) {
437 		if (amd_event_codes[i].pe_ev == pe) {
438 			config =
439 			    AMD_PMC_TO_EVENTMASK(amd_event_codes[i].pe_code);
440 			allowed_unitmask =
441 			    AMD_PMC_TO_UNITMASK(amd_event_codes[i].pe_mask);
442 			break;
443 		}
444 	}
445 	if (i == amd_event_codes_size)
446 		return (EINVAL);
447 
448 	unitmask = a->pm_md.pm_amd.pm_amd_config & AMD_PMC_UNITMASK;
449 	if ((unitmask & ~allowed_unitmask) != 0) /* disallow reserved bits */
450 		return (EINVAL);
451 
452 	if (unitmask && (caps & PMC_CAP_QUALIFIER) != 0)
453 		config |= unitmask;
454 
455 	if ((caps & PMC_CAP_THRESHOLD) != 0)
456 		config |= a->pm_md.pm_amd.pm_amd_config & AMD_PMC_COUNTERMASK;
457 
458 	/* Set at least one of the 'usr' or 'os' caps. */
459 	if ((caps & PMC_CAP_USER) != 0)
460 		config |= AMD_PMC_USR;
461 	if ((caps & PMC_CAP_SYSTEM) != 0)
462 		config |= AMD_PMC_OS;
463 	if ((caps & (PMC_CAP_USER | PMC_CAP_SYSTEM)) == 0)
464 		config |= (AMD_PMC_USR|AMD_PMC_OS);
465 
466 	if ((caps & PMC_CAP_EDGE) != 0)
467 		config |= AMD_PMC_EDGE;
468 	if ((caps & PMC_CAP_INVERT) != 0)
469 		config |= AMD_PMC_INVERT;
470 	if ((caps & PMC_CAP_INTERRUPT) != 0)
471 		config |= AMD_PMC_INT;
472 
473 	pm->pm_md.pm_amd.pm_amd_evsel = config; /* save config value */
474 
475 	PMCDBG2(MDP, ALL, 2, "amd-allocate ri=%d -> config=0x%x", ri, config);
476 
477 	return (0);
478 }
479 
480 /*
481  * Release machine dependent state associated with a PMC.  This is a
482  * no-op on this architecture.
483  */
484 static int
amd_release_pmc(int cpu,int ri,struct pmc * pmc __unused)485 amd_release_pmc(int cpu, int ri, struct pmc *pmc __unused)
486 {
487 	struct pmc_hw *phw __diagused;
488 
489 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
490 	    ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
491 	KASSERT(ri >= 0 && ri < amd_npmcs,
492 	    ("[amd,%d] illegal row-index %d", __LINE__, ri));
493 
494 	phw = &amd_pcpu[cpu]->pc_amdpmcs[ri];
495 
496 	KASSERT(phw->phw_pmc == NULL,
497 	    ("[amd,%d] PHW pmc %p non-NULL", __LINE__, phw->phw_pmc));
498 
499 	return (0);
500 }
501 
502 /*
503  * Start a PMC.
504  */
505 static int
amd_start_pmc(int cpu __diagused,int ri,struct pmc * pm)506 amd_start_pmc(int cpu __diagused, int ri, struct pmc *pm)
507 {
508 	const struct amd_descr *pd;
509 	uint64_t config;
510 
511 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
512 	    ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
513 	KASSERT(ri >= 0 && ri < amd_npmcs,
514 	    ("[amd,%d] illegal row-index %d", __LINE__, ri));
515 
516 	pd = &amd_pmcdesc[ri];
517 
518 	PMCDBG2(MDP, STA, 1, "amd-start cpu=%d ri=%d", cpu, ri);
519 
520 	/*
521 	 * Triggered by DF counters because all DF MSRs are shared.  We need to
522 	 * change the code to honor the per-package flag in the JSON event
523 	 * definitions.
524 	 */
525 	KASSERT(AMD_PMC_IS_STOPPED(pd->pm_evsel),
526 	    ("[amd,%d] pmc%d,cpu%d: Starting active PMC \"%s\"", __LINE__,
527 	    ri, cpu, pd->pm_descr.pd_name));
528 
529 	/* turn on the PMC ENABLE bit */
530 	config = pm->pm_md.pm_amd.pm_amd_evsel | AMD_PMC_ENABLE;
531 
532 	PMCDBG1(MDP, STA, 2, "amd-start config=0x%x", config);
533 
534 	wrmsr(pd->pm_evsel, config);
535 	return (0);
536 }
537 
538 /*
539  * Stop a PMC.
540  */
541 static int
amd_stop_pmc(int cpu __diagused,int ri,struct pmc * pm)542 amd_stop_pmc(int cpu __diagused, int ri, struct pmc *pm)
543 {
544 	const struct amd_descr *pd;
545 	uint64_t config;
546 	int i;
547 
548 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
549 	    ("[amd,%d] illegal CPU value %d", __LINE__, cpu));
550 	KASSERT(ri >= 0 && ri < amd_npmcs,
551 	    ("[amd,%d] illegal row-index %d", __LINE__, ri));
552 
553 	pd = &amd_pmcdesc[ri];
554 
555 	KASSERT(!AMD_PMC_IS_STOPPED(pd->pm_evsel),
556 	    ("[amd,%d] PMC%d, CPU%d \"%s\" already stopped",
557 		__LINE__, ri, cpu, pd->pm_descr.pd_name));
558 
559 	PMCDBG1(MDP, STO, 1, "amd-stop ri=%d", ri);
560 
561 	/* turn off the PMC ENABLE bit */
562 	config = pm->pm_md.pm_amd.pm_amd_evsel & ~AMD_PMC_ENABLE;
563 	wrmsr(pd->pm_evsel, config);
564 
565 	/*
566 	 * Due to NMI latency on newer AMD processors
567 	 * NMI interrupts are ignored, which leads to
568 	 * panic or messages based on kernel configuration
569 	 */
570 
571 	/* Wait for the count to be reset */
572 	for (i = 0; i < OVERFLOW_WAIT_COUNT; i++) {
573 		if (rdmsr(pd->pm_perfctr) & (1 << (pd->pm_descr.pd_width - 1)))
574 			break;
575 
576 		DELAY(1);
577 	}
578 
579 	return (0);
580 }
581 
582 /*
583  * Interrupt handler.  This function needs to return '1' if the
584  * interrupt was this CPU's PMCs or '0' otherwise.  It is not allowed
585  * to sleep or do anything a 'fast' interrupt handler is not allowed
586  * to do.
587  */
588 static int
amd_intr(struct trapframe * tf)589 amd_intr(struct trapframe *tf)
590 {
591 	struct amd_cpu *pac;
592 	struct pmc *pm;
593 	pmc_value_t v;
594 	uint64_t config, evsel, perfctr;
595 	uint32_t active = 0, count = 0;
596 	int i, error, retval, cpu;
597 
598 	cpu = curcpu;
599 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
600 	    ("[amd,%d] out of range CPU %d", __LINE__, cpu));
601 
602 	PMCDBG3(MDP, INT, 1, "cpu=%d tf=%p um=%d", cpu, tf, TRAPF_USERMODE(tf));
603 
604 	retval = 0;
605 
606 	pac = amd_pcpu[cpu];
607 
608 	retval = pmc_ibs_intr(tf);
609 	if (retval)
610 		goto done;
611 
612 	/*
613 	 * look for all PMCs that have interrupted:
614 	 * - look for a running, sampling PMC which has overflowed
615 	 *   and which has a valid 'struct pmc' association
616 	 *
617 	 * If found, we call a helper to process the interrupt.
618 	 *
619 	 * PMCs interrupting at the same time are collapsed into
620 	 * a single interrupt. Check all the valid pmcs for
621 	 * overflow.
622 	 */
623 	for (i = 0; i < amd_npmcs; i++) {
624 		if (amd_pmcdesc[i].pm_subclass != PMC_AMD_SUB_CLASS_CORE)
625 		    break;
626 
627 		if ((pm = pac->pc_amdpmcs[i].phw_pmc) == NULL ||
628 		    !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) {
629 			continue;
630 		}
631 
632 		/* Consider pmc with valid handle as active */
633 		active++;
634 
635 		if (!AMD_PMC_HAS_OVERFLOWED(i))
636 			continue;
637 
638 		retval = 1;	/* Found an interrupting PMC. */
639 
640 		if (pm->pm_state != PMC_STATE_RUNNING)
641 			continue;
642 
643 		/* Stop the PMC, reload count. */
644 		evsel   = amd_pmcdesc[i].pm_evsel;
645 		perfctr = amd_pmcdesc[i].pm_perfctr;
646 		v       = pm->pm_sc.pm_reloadcount;
647 		config  = rdmsr(evsel);
648 
649 		KASSERT((config & ~AMD_PMC_ENABLE) ==
650 		    (pm->pm_md.pm_amd.pm_amd_evsel & ~AMD_PMC_ENABLE),
651 		    ("[amd,%d] config mismatch reg=0x%jx pm=0x%jx", __LINE__,
652 			 (uintmax_t)config, (uintmax_t)pm->pm_md.pm_amd.pm_amd_evsel));
653 
654 		wrmsr(evsel, config & ~AMD_PMC_ENABLE);
655 		wrmsr(perfctr, AMD_RELOAD_COUNT_TO_PERFCTR_VALUE(v));
656 
657 		/* Restart the counter if logging succeeded. */
658 		error = pmc_process_interrupt(PMC_HR, pm, tf);
659 		if (error == 0)
660 			wrmsr(evsel, config);
661 	}
662 
663 	/*
664 	 * Due to NMI latency, there can be a scenario in which
665 	 * multiple pmcs gets serviced in an earlier NMI and we
666 	 * do not find an overflow in the subsequent NMI.
667 	 *
668 	 * For such cases we keep a per-cpu count of active NMIs
669 	 * and compare it with min(active pmcs, 2) to determine
670 	 * if this NMI was for a pmc overflow which was serviced
671 	 * in an earlier request or should be ignored.
672 	 */
673 	if (retval) {
674 		DPCPU_SET(nmi_counter, min(2, active));
675 	} else {
676 		if ((count = DPCPU_GET(nmi_counter))) {
677 			retval = 1;
678 			DPCPU_SET(nmi_counter, --count);
679 		}
680 	}
681 
682 done:
683 	if (retval)
684 		counter_u64_add(pmc_stats.pm_intr_processed, 1);
685 	else
686 		counter_u64_add(pmc_stats.pm_intr_ignored, 1);
687 
688 	PMCDBG1(MDP, INT, 2, "retval=%d", retval);
689 	return (retval);
690 }
691 
692 /*
693  * Describe a PMC.
694  */
695 static int
amd_describe(int cpu,int ri,struct pmc_info * pi,struct pmc ** ppmc)696 amd_describe(int cpu, int ri, struct pmc_info *pi, struct pmc **ppmc)
697 {
698 	const struct amd_descr *pd;
699 	struct pmc_hw *phw;
700 
701 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
702 	    ("[amd,%d] illegal CPU %d", __LINE__, cpu));
703 	KASSERT(ri >= 0 && ri < amd_npmcs,
704 	    ("[amd,%d] row-index %d out of range", __LINE__, ri));
705 
706 	phw = &amd_pcpu[cpu]->pc_amdpmcs[ri];
707 	pd  = &amd_pmcdesc[ri];
708 
709 	strlcpy(pi->pm_name, pd->pm_descr.pd_name, sizeof(pi->pm_name));
710 	pi->pm_class = pd->pm_descr.pd_class;
711 
712 	if ((phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) != 0) {
713 		pi->pm_enabled = true;
714 		*ppmc          = phw->phw_pmc;
715 	} else {
716 		pi->pm_enabled = false;
717 		*ppmc          = NULL;
718 	}
719 
720 	return (0);
721 }
722 
723 /*
724  * Return the MSR address of the given PMC.
725  */
726 static int
amd_get_msr(int ri,uint32_t * msr)727 amd_get_msr(int ri, uint32_t *msr)
728 {
729 	int df_idx;
730 
731 	KASSERT(ri >= 0 && ri < amd_npmcs,
732 	    ("[amd,%d] ri %d out of range", __LINE__, ri));
733 
734 	/*
735 	 * Map counter row index to RDPMC ECX value.
736 	 *
737 	 * AMD BKDG 24594 rev 3.37, page 440,
738 	 * "RDPMC Read Performance-Monitoring Counter":
739 	 *   ECX 0-5:   Core counters 0-5
740 	 *   ECX 6-9:   DF/Northbridge counters 0-3
741 	 *   ECX 10-15: L3 Cache counters 0-5
742 	 *   ECX 16-27: DF/Northbridge counters 4-15
743 	 *
744 	 * AMD PPR 57930-A0 section 2.1.9,
745 	 * "Register Sharing" for DF counter details.
746 	 */
747 	if (ri < amd_core_npmcs) {
748 		/* ECX 0-5: Core counters */
749 		*msr = ri;
750 	} else if (ri < amd_core_npmcs + amd_l3_npmcs) {
751 		/* ECX 10-15: L3 Cache counters */
752 		*msr = 10 + (ri - amd_core_npmcs);
753 	} else {
754 		/* ECX 6-9: DF counters 0-3
755 		 * ECX 16-27: DF counters 4-15 */
756 		df_idx = ri - amd_core_npmcs - amd_l3_npmcs;
757 		if (df_idx < 4)
758 			*msr = 6 + df_idx;
759 		else if (df_idx < 16)
760 			*msr = 16 + (df_idx - 4);
761 		else
762 			return (EINVAL);
763 	}
764 	return (0);
765 }
766 
767 /*
768  * Return the capabilities of the given PMC.
769  */
770 static int
amd_get_caps(int ri,uint32_t * caps)771 amd_get_caps(int ri, uint32_t *caps)
772 {
773 	KASSERT(ri >= 0 && ri < amd_npmcs,
774 	    ("[amd,%d] ri %d out of range", __LINE__, ri));
775 
776 	*caps = amd_pmcdesc[ri].pm_descr.pd_caps;
777 
778 	return (0);
779 }
780 
781 /*
782  * Processor-dependent initialization.
783  */
784 static int
amd_pcpu_init(struct pmc_mdep * md,int cpu)785 amd_pcpu_init(struct pmc_mdep *md, int cpu)
786 {
787 	struct amd_cpu *pac;
788 	struct pmc_cpu *pc;
789 	struct pmc_hw  *phw;
790 	int first_ri, n;
791 
792 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
793 	    ("[amd,%d] insane cpu number %d", __LINE__, cpu));
794 
795 	PMCDBG1(MDP, INI, 1, "amd-init cpu=%d", cpu);
796 
797 	amd_pcpu[cpu] = pac = malloc(sizeof(struct amd_cpu), M_PMC,
798 	    M_WAITOK | M_ZERO);
799 
800 	/*
801 	 * Set the content of the hardware descriptors to a known
802 	 * state and initialize pointers in the MI per-cpu descriptor.
803 	 */
804 	pc = pmc_pcpu[cpu];
805 	first_ri = md->pmd_classdep[PMC_MDEP_CLASS_INDEX_K8].pcd_ri;
806 
807 	KASSERT(pc != NULL, ("[amd,%d] NULL per-cpu pointer", __LINE__));
808 
809 	for (n = 0, phw = pac->pc_amdpmcs; n < amd_npmcs; n++, phw++) {
810 		phw->phw_state = PMC_PHW_FLAG_IS_ENABLED |
811 		    PMC_PHW_CPU_TO_STATE(cpu) | PMC_PHW_INDEX_TO_STATE(n);
812 		phw->phw_pmc = NULL;
813 		pc->pc_hwpmcs[n + first_ri] = phw;
814 	}
815 
816 	return (0);
817 }
818 
819 /*
820  * Processor-dependent cleanup prior to the KLD being unloaded.
821  */
822 static int
amd_pcpu_fini(struct pmc_mdep * md,int cpu)823 amd_pcpu_fini(struct pmc_mdep *md, int cpu)
824 {
825 	struct amd_cpu *pac;
826 	struct pmc_cpu *pc;
827 	int first_ri, i;
828 
829 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
830 	    ("[amd,%d] insane cpu number (%d)", __LINE__, cpu));
831 
832 	PMCDBG1(MDP, INI, 1, "amd-cleanup cpu=%d", cpu);
833 
834 	/*
835 	 * Next, free up allocated space.
836 	 */
837 	if ((pac = amd_pcpu[cpu]) == NULL)
838 		return (0);
839 
840 	amd_pcpu[cpu] = NULL;
841 
842 #ifdef	HWPMC_DEBUG
843 	for (i = 0; i < AMD_NPMCS_K8; i++) {
844 		KASSERT(pac->pc_amdpmcs[i].phw_pmc == NULL,
845 		    ("[amd,%d] CPU%d/PMC%d in use", __LINE__, cpu, i));
846 		KASSERT(AMD_PMC_IS_STOPPED(AMD_PMC_EVSEL_0 + i),
847 		    ("[amd,%d] CPU%d/PMC%d not stopped", __LINE__, cpu, i));
848 	}
849 #endif
850 
851 	pc = pmc_pcpu[cpu];
852 	KASSERT(pc != NULL, ("[amd,%d] NULL per-cpu state", __LINE__));
853 
854 	first_ri = md->pmd_classdep[PMC_MDEP_CLASS_INDEX_K8].pcd_ri;
855 
856 	/*
857 	 * Reset pointers in the MI 'per-cpu' state.
858 	 */
859 	for (i = 0; i < amd_npmcs; i++)
860 		pc->pc_hwpmcs[i + first_ri] = NULL;
861 
862 	free(pac, M_PMC);
863 	return (0);
864 }
865 
866 /*
867  * Initialize ourselves.
868  */
869 struct pmc_mdep *
pmc_amd_initialize(void)870 pmc_amd_initialize(void)
871 {
872 	struct pmc_classdep *pcd;
873 	struct pmc_mdep *pmc_mdep;
874 	enum pmc_cputype cputype;
875 	int error, i, ncpus, nclasses;
876 	int family, model, stepping;
877 	struct amd_descr *d;
878 
879 	/*
880 	 * The presence of hardware performance counters on the AMD
881 	 * Athlon, Duron or later processors, is _not_ indicated by
882 	 * any of the processor feature flags set by the 'CPUID'
883 	 * instruction, so we only check the 'instruction family'
884 	 * field returned by CPUID for instruction family >= 6.
885 	 */
886 
887 	family = CPUID_TO_FAMILY(cpu_id);
888 	model = CPUID_TO_MODEL(cpu_id);
889 	stepping = CPUID_TO_STEPPING(cpu_id);
890 
891 	if (family == 0x18)
892 		snprintf(pmc_cpuid, sizeof(pmc_cpuid), "HygonGenuine-%d-%02X-%X",
893 		    family, model, stepping);
894 	else
895 		snprintf(pmc_cpuid, sizeof(pmc_cpuid), "AuthenticAMD-%d-%02X-%X",
896 		    family, model, stepping);
897 
898 	switch (cpu_id & 0xF00) {
899 	case 0xF00:		/* Athlon64/Opteron processor */
900 		cputype = PMC_CPU_AMD_K8;
901 		break;
902 	default:
903 		printf("pmc: Unknown AMD CPU %x %d-%d.\n", cpu_id, family,
904 		    model);
905 		return (NULL);
906 	}
907 
908 	/*
909 	 * From PPR for AMD Family 1Ah, a new cpuid leaf specifies the maximum
910 	 * number of PMCs of each type.  If we do not have that leaf, we use
911 	 * the prior default values that are only valid if we have the feature
912 	 * bit enabled in CPU.
913 	 */
914 	if ((amd_feature2 & AMDID2_PCXC) != 0) {
915 		amd_core_npmcs = AMD_PMC_CORE_DEFAULT;
916 	} else {
917 		amd_core_npmcs = AMD_NPMCS_K8;
918 	}
919 	amd_l3_npmcs = AMD_PMC_L3_DEFAULT;
920 	amd_df_npmcs = AMD_PMC_DF_DEFAULT;
921 
922 	if (cpu_exthigh >= CPUID_EXTPERFMON) {
923 		u_int regs[4];
924 		do_cpuid(CPUID_EXTPERFMON, regs);
925 		if (regs[1] != 0) {
926 			amd_core_npmcs = EXTPERFMON_CORE_PMCS(regs[1]);
927 			amd_df_npmcs = EXTPERFMON_DF_PMCS(regs[1]);
928 		}
929 	}
930 
931 	/* Enable the newer core counters */
932 	for (i = 0; i < amd_core_npmcs; i++) {
933 		d = &amd_pmcdesc[i];
934 		snprintf(d->pm_descr.pd_name, PMC_NAME_MAX,
935 		    "K8-%d", i);
936 		d->pm_descr.pd_class = PMC_CLASS_K8;
937 		d->pm_descr.pd_caps = AMD_PMC_CAPS;
938 		/*
939 		 * Zen 5 can precisely count retire events.
940 		 *
941 		 * Refer to PPR Vol 1 for AMD Family 1Ah Model 02h C1 57238
942 		 * Rev. 0.24 September 29, 2024.
943 		 */
944 		if ((family >= 0x1a) && (i == 2))
945 			d->pm_descr.pd_caps |= PMC_CAP_PRECISE;
946 		d->pm_descr.pd_width = 48;
947 		if ((amd_feature2 & AMDID2_PCXC) != 0) {
948 			d->pm_evsel = AMD_PMC_CORE_BASE + 2 * i;
949 			d->pm_perfctr = AMD_PMC_CORE_BASE + 2 * i + 1;
950 		} else {
951 			d->pm_evsel = AMD_PMC_EVSEL_0 + i;
952 			d->pm_perfctr = AMD_PMC_PERFCTR_0 + i;
953 		}
954 		d->pm_subclass = PMC_AMD_SUB_CLASS_CORE;
955 	}
956 	amd_npmcs = amd_core_npmcs;
957 
958 	if ((amd_feature2 & AMDID2_PTSCEL2I) != 0) {
959 		/* Enable the LLC/L3 counters */
960 		for (i = 0; i < amd_l3_npmcs; i++) {
961 			d = &amd_pmcdesc[amd_npmcs + i];
962 			snprintf(d->pm_descr.pd_name, PMC_NAME_MAX,
963 			    "K8-L3-%d", i);
964 			d->pm_descr.pd_class = PMC_CLASS_K8;
965 			d->pm_descr.pd_caps = AMD_PMC_L3_CAPS;
966 			d->pm_descr.pd_width = 48;
967 			d->pm_evsel = AMD_PMC_L3_BASE + 2 * i;
968 			d->pm_perfctr = AMD_PMC_L3_BASE + 2 * i + 1;
969 			d->pm_subclass = PMC_AMD_SUB_CLASS_L3_CACHE;
970 		}
971 		amd_npmcs += amd_l3_npmcs;
972 	}
973 
974 	if ((amd_feature2 & AMDID2_PNXC) != 0) {
975 		/* Enable the data fabric counters */
976 		for (i = 0; i < amd_df_npmcs; i++) {
977 			d = &amd_pmcdesc[amd_npmcs + i];
978 			snprintf(d->pm_descr.pd_name, PMC_NAME_MAX,
979 			    "K8-DF-%d", i);
980 			d->pm_descr.pd_class = PMC_CLASS_K8;
981 			d->pm_descr.pd_caps = AMD_PMC_DF_CAPS;
982 			d->pm_descr.pd_width = 48;
983 			d->pm_evsel = AMD_PMC_DF_BASE + 2 * i;
984 			d->pm_perfctr = AMD_PMC_DF_BASE + 2 * i + 1;
985 			d->pm_subclass = PMC_AMD_SUB_CLASS_DATA_FABRIC;
986 		}
987 		amd_npmcs += amd_df_npmcs;
988 	}
989 
990 	/*
991 	 * Allocate space for pointers to PMC HW descriptors and for
992 	 * the MDEP structure used by MI code.
993 	 */
994 	amd_pcpu = malloc(sizeof(struct amd_cpu *) * pmc_cpu_max(), M_PMC,
995 	    M_WAITOK | M_ZERO);
996 
997 	/*
998 	 * These processors have two or three classes of PMCs: the TSC,
999 	 * programmable PMCs, and AMD IBS.
1000 	 */
1001 	if ((amd_feature2 & AMDID2_IBS) != 0) {
1002 		nclasses = 3;
1003 	} else {
1004 		nclasses = 2;
1005 	}
1006 
1007 	pmc_mdep = pmc_mdep_alloc(nclasses);
1008 
1009 	ncpus = pmc_cpu_max();
1010 
1011 	/* Initialize the TSC. */
1012 	error = pmc_tsc_initialize(pmc_mdep, ncpus);
1013 	if (error != 0)
1014 		goto error;
1015 
1016 	/* Initialize AMD K8 PMC handling. */
1017 	pcd = &pmc_mdep->pmd_classdep[PMC_MDEP_CLASS_INDEX_K8];
1018 
1019 	pcd->pcd_caps		= AMD_PMC_CAPS;
1020 	pcd->pcd_class		= PMC_CLASS_K8;
1021 	pcd->pcd_num		= amd_npmcs;
1022 	pcd->pcd_ri		= pmc_mdep->pmd_npmc;
1023 	pcd->pcd_width		= 48;
1024 
1025 	pcd->pcd_allocate_pmc	= amd_allocate_pmc;
1026 	pcd->pcd_config_pmc	= amd_config_pmc;
1027 	pcd->pcd_describe	= amd_describe;
1028 	pcd->pcd_get_config	= amd_get_config;
1029 	pcd->pcd_get_msr	= amd_get_msr;
1030 	pcd->pcd_pcpu_fini	= amd_pcpu_fini;
1031 	pcd->pcd_pcpu_init	= amd_pcpu_init;
1032 	pcd->pcd_read_pmc	= amd_read_pmc;
1033 	pcd->pcd_release_pmc	= amd_release_pmc;
1034 	pcd->pcd_start_pmc	= amd_start_pmc;
1035 	pcd->pcd_stop_pmc	= amd_stop_pmc;
1036 	pcd->pcd_write_pmc	= amd_write_pmc;
1037 	pcd->pcd_get_caps	= amd_get_caps;
1038 
1039 	pmc_mdep->pmd_cputype	= cputype;
1040 	pmc_mdep->pmd_intr	= amd_intr;
1041 	pmc_mdep->pmd_switch_in	= amd_switch_in;
1042 	pmc_mdep->pmd_switch_out = amd_switch_out;
1043 
1044 	pmc_mdep->pmd_npmc	+= amd_npmcs;
1045 
1046 	amd_init_policy();
1047 
1048 	PMCDBG0(MDP, INI, 0, "amd-initialize");
1049 
1050 	if (nclasses >= 3) {
1051 		error = pmc_ibs_initialize(pmc_mdep, ncpus);
1052 		if (error != 0)
1053 			goto error;
1054 	}
1055 
1056 	return (pmc_mdep);
1057 
1058 error:
1059 	free(pmc_mdep, M_PMC);
1060 	return (NULL);
1061 }
1062 
1063 /*
1064  * Finalization code for AMD CPUs.
1065  */
1066 void
pmc_amd_finalize(struct pmc_mdep * md)1067 pmc_amd_finalize(struct pmc_mdep *md)
1068 {
1069 	PMCDBG0(MDP, INI, 1, "amd-finalize");
1070 
1071 	pmc_tsc_finalize(md);
1072 
1073 	for (int i = 0; i < pmc_cpu_max(); i++)
1074 		KASSERT(amd_pcpu[i] == NULL,
1075 		    ("[amd,%d] non-null pcpu cpu %d", __LINE__, i));
1076 
1077 	free(amd_pcpu, M_PMC);
1078 	amd_pcpu = NULL;
1079 }
1080